<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2918">
  <Title>Using Gazetteers in Discriminative Information Extraction</Title>
  <Section position="11" start_page="578" end_page="578" type="concl">
    <SectionTitle>
9 Conclusion and future work
</SectionTitle>
    <Paragraph position="0"> In this paper we have identi ed and analysed negative effects that can be introduced to maxent-style models by the inclusion of highly discriminative gazetteer features. We have shown that such effects manifest themselves through errors that generally result from the model's over-dependence on the gazetteer features for decision making. To overcome this problem a more careful treatment of these features is required during training. We have proposed a solution that involves quarantining the features and training them separately to the other features in the model, then decoding the separate models with a logarithmic opinion pool. In fact, the LOP provides a natural way to handle the problem, with different constituent models for the different feature types. The method leads to much greater accuracy, and allows the power of gazetteer features to be more effectively harnessed. Finally, we have identi ed other feature sets with gazetteer feature-like properties and shown that similar results may be obtained using our method with these feature sets.</Paragraph>
    <Paragraph position="1"> In this paper we de ned intuitively-motivated feature partitions (gazetteer feature-based or otherwise) using heuristics. In future work we will focus on automatically determining such partitions.</Paragraph>
  </Section>
class="xml-element"></Paper>